- Added the `xd://` virtual device protocol (`internal-urls/xd-protocol.ts`, `tools/xdev.ts`): tools declaring `loadMode: "discoverable"` are unmounted from the request tools array and driven via `read xd://` (list/docs+schema) and `write xd://<tool>` (execute), gated by the `tools.xdev` setting (default on) and inlined into the system prompt.
- Merged the `irc`, `job`, and `launch` tools into a single `hub` tool (`tools/hub/`, `async/job-manager.ts`): messaging keeps `send`/`inbox`/`list`, job control maps to `wait`/`cancel`/`jobs`, process supervision keeps `start`/`logs`/`stop`/`restart`/`describe` with `ps`, and the unified `wait` races background jobs against peer messages; SDK `IrcTool`/`JobTool`/`LaunchTool` are replaced by `HubTool`.
- Removed the hidden `resolve` tool in favor of the `xd://resolve`/`xd://reject`/`xd://propose` resolution devices, auto-including `write` whenever a deferrable tool or plan mode is present.
- Removed the BM25 tool-discovery system: the `search_tool_bm25` tool, the `tool-discovery` module, the `tools.discoveryMode`/`mcp.discoveryMode`/`mcp.discoveryDefaultServers`/`tools.essentialOverride` settings, per-tool MCP selection, and the `mcp_tool_selection` message type.
- Unified tool presentation on `ToolLoadMode` (`essential`|`discoverable`), replacing the custom-tool `xdev?: boolean` opt-out; custom, extension, MCP, RPC host, image-generation, and TTS tools now default to `discoverable`, and added a `satisfies` predicate to `SoftToolRequirement`.
- Removed the standalone `ssh` command tool and `ssh/ssh-executor` (the `ssh://` read/write/search protocol stays), and made `--tools` address hidden built-ins.
- Updated collab-web to render `xd://` dispatches and `hub` op families, dropped the `search_tool_bm25`/`ssh`/`report-finding` renderers, refreshed tool docs and prompts, and migrated the affected tests and changelogs.
Registered ACP session disposal with postmortem and replaced the hard EOF exit with the awaited graceful shutdown path.
Classified stdio-write EPIPE separately from worker IPC EPIPE so ACP peer loss exits successfully after cleanup.
Fixes#4788
- Normalized completed image_generation_call results into assistant image blocks.
- Persisted image bytes through the session blob store and rendered them in live, replay, ACP, proxy, telemetry, and HTML paths.
- Added response normalization, persistence, and TUI rendering regressions.
Fixes#4768
- Added `isAcceptedElicitation` in the ACP agent to narrow accepted elicitations before accessing response content.
- Added `isFormElicitation` in ACP tests and used it to narrow form-mode requests before assertions.
- Added a fallback to emit error messages during `agent_end` if no error was previously streamed during the turn.
- Added tracking to prevent duplicate error messages when a provider error is successfully delivered during streaming.
- Added a stderr hint for interactive users launching the ACP server directly from a terminal.
The assistant message_end fan-out is fire-and-forget in the session layer
and can be parked on extension delivery while agent_end is flushed through
#endInFlight, so agent_end can overtake it. #finishPrompt then unsubscribes
the prompt turn and the mapAssistantMessageEnd fallback never runs: an ACP
client that only received agent_thought_chunk updates (thinking streamed,
text arrived only on the trailing message) stays stuck on the thinking
block with no visible answer. On agent_end, emit the last assistant
message's text before resolving the prompt when live-message progress shows
no text was ever delivered, and defer the live-state reset past that flush
so a late message_end cannot resurrect fresh progress and double-emit.
Fixes#4902
Extensions calling ctx.ui.addAutocompleteProvider (e.g. @ff-labs/pi-fff)
crashed at load with 'TypeError: ... is not a function' because omp's
ExtensionAPI.ui omitted pi's autocomplete-provider API; the throw also
aborted the rest of a try/catch-guarded session_start init.
ExtensionUIContext now declares addAutocompleteProvider(factory).
Interactive mode stacks each factory on the built-in editor provider in
registration order, re-applies the stack on every slash-command refresh,
and skips throwing/malformed factories; RPC, ACP, and headless contexts
accept the factory as a no-op, matching upstream pi's RPC behavior.
Fixes#4919
Extension sendUserMessage() without deliverAs fell through to prompt(),
which throws AgentBusyError during an active stream; the message was
dropped and surfaced as 'Extension sendUserMessage failed'. Route the
omitted-deliverAs path through prompt() with streamingBehavior 'steer'
so streaming queues a steer with normal prompt-flow side effects
(keyword notices, advisor auto-resume reset) and idle still starts a
turn.
ACP skill-command prompts now pass streamingBehavior 'steer'; the RPC
skill fast-path honors the prompt command's streamingBehavior field
(default steer) like the plain-prompt path already did. Documented the
extension-facing delivery semantics.
Synthesized from PR #4942 (prompt-flow steer routing, docs, tests) and
PR #4922 (RPC streamingBehavior threading, steer regression test);
dropped PR #4942's unrelated workflow-notice.md ellipsis churn.
Fixes#4923
Co-authored-by: roboomp <omp@can.ac>
Co-authored-by: metaphorics <metaphorics@users.noreply.github.com>
- Added a Usage.orchestration sidecar for provider-side service tokens so Responses/Codex totals and costs stay accurate without inflating visible prompt input/cache buckets.
- Updated Codex/WebSocket usage, session/status aggregates, and usage reporting to preserve orchestration-aware totals.
- Added regressions for OpenAI Responses accounting, Codex WebSocket terminal usage, cost calculation, and session aggregation.
Fixes#4469
The mid-prompt slash skill autocomplete added in #3654 replaced the
entire editor draft with /skill:<name> on accept so the dispatcher
(which only matched leading /skill:) would still fire. That wiped
every keystroke the user had typed before reaching for the skill.
Insert the /skill:<name> token at the cursor in the TUI editor —
replacing only the partial /sk slash token, leaving prose before and
after intact — and extend the skill-command parser so a /skill:<name>
token surrounded by whitespace is recognized as an invocation too,
with the surrounding prose threaded through to the skill as args.
The parser change is shared across all three dispatch sites
(interactive TUI, ACP, RPC) via a new parseSkillInvocation helper
in extensibility/skills, so the three Map<string,string> /
session.skills lookups stay aligned on the same parse.
Fixes#3913
User-invoked skills (typed /skill:, steered, follow-up, interrupted/
resumed via compaction, ACP, RPC) only appended a bare "Skill: <path>"
line, so the model neither learned the user had invoked that specific
skill nor where the skill directory was. Relative paths in skill bodies
(scripts/, templates/) could not be resolved.
Route all user-invoked paths through a self-identifying, baseDir-aware
prompt template; keep hidden autoload skills on the minimal non-user
format. Interactive skillCommands now carries the loaded Skill object
instead of a bare path so baseDir flows through without reconstruction.
The invocation kind defaults to "user" to keep buildSkillPromptMessage
source-compatible.
Op: correct
Restores: spec:user-invoked-skill-prompt-self-identifies-and-exposes-skill-directory
- Migrated 288 lines of scattered error classification logic from `utils/error-id.ts` into a cohesive `packages/ai/src/error/` module with 13 specialized submodules covering flags, classes, OAuth, providers, rate-limiting, and finalization.
- Replaced 100+ generic `Error` throws across 60+ provider and registry files with semantic `AIError.*` classes (e.g., `AIError.MissingApiKeyError`, `AIError.OAuthError`, `AIError.ProviderResponseError`), improving error diagnostics and retry logic.
- Consolidated error utility imports from `pi-utils` and scattered classification functions into a single `AIError` namespace, reducing coupling and simplifying error handling across all packages.
Resolved live ACP generate_image payloads through the blob store before emitting image content while keeping rawOutput compact.
Added regression coverage for content[] image blocks and details.images entries without duplicating blob refs as fallback text.
Fixes#3623
- Renamed the `find` and `search` tools to `glob` and `grep` respectively across the codebase to improve command clarity.
- Implemented full-stack support for the renamed tools, including CLI arguments, system prompts, SDK exports, and tool registration.
- Added automated migration logic in `settings` to transform legacy `find` and `search` configuration keys to their new equivalents.
- Updated the `collab-web` renderer registry to ensure backwards compatibility with legacy tool outputs.
- Transitioned the eval tool from batch multi-cell execution to a single-step input structure with flat parameters.
- Updated core agent logic, UI components, and documentation to support state persistence across incremental eval calls.
- Restricted bash tool capabilities by requiring explicit use of `read` or `find` instead of `ls` or `find`.
- Added support for Ruby and Julia language runtimes to the eval tool and associated web renderers.
Added the ACP mobile speech.models.list method so voice settings can fetch static local STT, TTS, and voice catalog options without invoking setup/download paths.
Fixes#3011
- Passed USER_INTERRUPT_LABEL through abort paths in collab, ACP, RPC, runtime, and SDK flows.
- Added userInitiated to synthetic continue inputs and session prompt calls.
- Suppressed advisor auto-resume during user aborts and preserved queued concerns.
- Cleared suppression on user prompts and reclaimed parked advisor cards on abort settle.
- Canonicalized assistant and thinking messages by trimming and collapsing dot text.
- Skipped rendering assistant and thinking blocks when canonicalized content was empty.
- Filtered ACP thinking notifications and session outputs to ignore placeholder content.
- Added canonicalizeMessage tests for undefined, blank, whitespace, and dot-only inputs.
- Added session-domain modules and exports for session-entries, context, listing, loader, and migrations.
- Changed persistence to async append writes plus writeTextAtomic, removing sync line APIs.
- Added compaction-aware session context rebuild with dangling tool-call cleanup.
- Added resumable session resolution with status inference, id/stem/suffix matching, and backup recovery.
- Filtered dot-only or blank thinking blocks so they no longer render as assistant thought.
- Adjusted assistant-message and streaming-reveal logic to use visible-thinking helpers for consistency.
- Removed `userMessageId` from `AcpAgent` prompt state and response payloads.
- Removed `models` from new/load/resume/fork session responses and deleted model-state building.
- Removed `unstable_setSessionModel` and routed model changes through `setSessionConfigOption`.
parseSlashCommand treats ':' as a name/args separator, so an extension
command like 'model:foo' was advertised in available_commands_update but
dispatched to the '/model' builtin. Filter such names via
isAcpBuiltinShadowedName, and fix the FakeAgentSession prompt stubs in
acp-agent.test.ts to return true now that AgentSession.prompt() reports
whether the agent was invoked (6 tests were failing against the new
early-finish path).
Addresses review feedback on #2052.
Extension commands (e.g. /sonnet) and TypeScript custom commands that
consume the input without calling the LLM return early from
session.prompt() with no agent turn. In ACP mode this left the pending
prompt promise unresolved, hanging the client forever.
Change session.prompt() to return Promise<boolean>: true when the LLM
was invoked, false when the command was fully handled locally.
#runPromptOrCommand calls #finishPrompt immediately on a false return so
the ACP turn completes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ACP_BUILTIN_SLASH_COMMANDS only carries primary names; the reserved set
passed to getRegisteredCommands was therefore missing aliases like
"models" (/model) and "force:" (/force). An extension registering one
of these aliases would appear in the palette but the builtin would win
at dispatch time (lookupBuiltinSlashCommand searches aliases too).
Export ACP_BUILTIN_RESERVED_NAMES from acp-builtins — the union of all
primary names and aliases for ACP-surfaced builtins — and use it as the
reserved set. Widen getRegisteredCommands parameter to ReadonlySet since
it only calls .has().
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Dispatch in AgentSession runs #tryExecuteExtensionCommand before
#tryExecuteCustomCommand, so the palette must reflect the same order.
Moving the extension-runner block before session.customCommands ensures
that on a name collision the advertised command matches what will
actually execute.
Update the test to assert the extension description wins over the
colliding custom TS description.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Update ordering comment in #buildAvailableCommands to document the
extension tier and explain why skills/custom TS commands intentionally
shadow extension commands (unlike interactive mode)
- Add CHANGELOG entry under [Unreleased]
- Add regression test: verifies extension commands surface in
available_commands_update and that a builtin-colliding extension
command is excluded via the reserved-set, with no duplicates
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
`#buildAvailableCommands` was omitting commands registered by extensions
via `extensionRunner.getRegisteredCommands()`. ACP clients (e.g. Zed)
never saw these in the `available_commands_update` notification, so they
couldn't forward the corresponding slash commands to the agent.
Mirrors the interactive-mode pattern: pass ACP builtin names as the
reserved set so extensions cannot shadow core commands.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The implicit-cancel overlap path dropped the cleanup promise, so an
abort() that hung past the cleanup timeout left the managed session
registered with a still-streaming AgentSession — the explicit cancel()
path closes it in that case. Mirror that handling and cover it with a
regression test.
Addresses review feedback on #2186.
When the user presses Stop in Zed and immediately types a new message,
the new session/prompt RPC can arrive before (or without) a preceding
session/cancel notification. The previous guard threw an error in that
case, leaving the session stuck and blocking further interaction.
Replace the throw with an implicit cancel: call #beginCancelCleanup on
the unsettled turn so it resolves with stopReason:"cancelled", then let
#queuePrompt serialize the new prompt behind the abort cleanup as it
already does when session/cancel is called explicitly. #beginCancelCleanup
is idempotent so a concurrent explicit cancel notification is a no-op.
Updated the test to assert the new contract: overlapping prompt
auto-cancels the first turn and the second is processed normally.
- Replaced approved-plan renaming with `resolveApprovedPlan` resolution and state/slug lookup.
- Updated ACP and interactive apply flows to propagate canonical `planFilePath` instead of renamed paths.
- Added local plan fallback lookup by mtime for unresolved slugs after plan approval.
- Restricted plan-mode writes to `local://` plan artifacts and simplified path handling.
#requestAcpPlanApprovalChoice returned 'value !== REFINE_OPTION' which
treats any non-refine outcome — explicit dismissal, transport failure,
abort, or timeout — as approval. That meant closing the elicitation
dialog (or the request failing) granted the agent write access without
explicit user consent.
Tighten to 'value === APPROVE_OPTION' so only the explicit approve
selection passes the gate. Refine, dismissal, and every other outcome
fall through to refine semantics: the caller keeps plan mode active and
returns guidance text instead of renaming the plan and exiting.
Adds a regression test that drives a form-capable harness with a
cancelled elicitation and asserts the plan file is preserved, plan
mode/handler stay active, plan reference stays unset, and no mode-exit
notifications are emitted.
ACP plan approval exited plan mode internally and emitted the current-mode
notification, but it did not push the matching mode config-option update.
Clients that cache or render the config selector could therefore keep
showing mode=plan after approval switched the session back to default.
The approval path now emits the same config_option_update used by other
mode transitions, and the regression test asserts that the approval-driven
exit updates both current_mode_update and the mode config option.
ACP plan mode set the plan-mode state but never installed the standing
resolve handler that interactive mode wires in #enterPlanMode. The agent
in plan mode dutifully called resolve { action: 'apply', extra.title }
to submit its plan; ResolveTool.execute then consulted
peekQueueInvoker() ?? peekStandingResolveHandler(), found neither, and
threw 'No pending action to resolve. Nothing to apply or discard.' —
stranding the agent with no path out of plan mode in Zed (and any other
ACP client).
#applyModeChange now registers a standing handler when entering
mode: 'plan' and clears it when leaving. The handler:
- validates the plan file exists at the configured '/data/workspaces/can1357__oh-my-pi__1869/.omp-session/2026-06-04T15-12-15-302Z_019e9331-31c6-7000-9bce-6be119505a1d/local' path,
- normalizes the agent-supplied title to a safe filename stem,
- asks the ACP client to confirm via unstable_createElicitation when
the client advertises elicitation.form (auto-approves otherwise so
the agent never gets stuck on a client without the surface),
- renames the plan to '/data/workspaces/can1357__oh-my-pi__1869/.omp-session/2026-06-04T15-12-15-302Z_019e9331-31c6-7000-9bce-6be119505a1d/local/<title>.md,'
- sets the plan reference path so the next turn injects the plan as
context, then clears the standing handler + plan-mode state and
emits current_mode_update so the client UI reflects the exit.
Refinement (user picks 'Refine plan') returns guidance text and leaves
plan mode active so the agent can iterate.
Fixes#1869
- Renamed `TodoWriteTool` to `TodoTool` and its source/prompt files.
- Updated tool registration, schema, renderers, and gating to `todo`.
- Adjusted cursor provider native tool names and tests to match.
- Renamed strike-animation constants and `todo-error-reminder` type.
When Zed provisions MCP servers through `session/new.mcpServers`, the
ACP agent
correctly connects to them and registers their tools via
`refreshMCPTools`, but
the tools were never activated. `refreshMCPTools` builds the next
active tool
set from `getSelectedMCPToolNames()`, which — with discovery disabled —
returns
only MCP tools already in the active set. Since ACP sessions start with
no MCP
tools, this created a circular deadlock: tools could only become active
if they
were already active.
Added an optional `{ activateAll?: boolean }` parameter to
`refreshMCPTools`.
When true, every newly registered tool is force-activated regardless of
prior
selection. `AcpAgent#configureMcpServers` passes `activateAll: true` on
both
call sites so client-provisioned tools are immediately usable.
- Changed tiny-device preference resolution to always default to CPU instead of platform-specific DirectML/CUDA heuristics.
- Updated settings schema, documentation, and changelog text to describe the CPU default while keeping accelerated providers behind explicit `providers.tinyModelDevice`/`PI_TINY_DEVICE` choices.
- Updated ACP thought-level resolution to use a helper that prefers session.configuredThinkingLevel and falls back to session.thinkingLevel when unavailable.
- Added a private AcpAgent method to safely compute configured thinking level without assuming a method exists on every session.
- Extended the ACP event-mapper replay test session with configuredThinkingLevel so it follows the new lookup path.
can1357